Appendix B. Integration
and Survey Methods
T
he U.S. Bureau of Labor Statistics (BLS or the Bureau)
has been gathering information on the spending patterns
and living costs of American consumers for more than a century—since the first such survey in 1888–91. Survey methods have been improved and refined over the years. A major
methodological improvement, first used in the 1972–73 survey, was the introduction of two separate surveys—a quarterly interview survey and a weekly diary or recordkeeping
survey—rather than a single interview survey, relying primarily on annual recall by survey participants. The Bureau added
a further enhancement in 1980, when it began data collection
for the survey on a continuing basis, rather than at intervals of
about 10 years.
The Bureau designed the surveys so that each has its own
questionnaire and sample. For the quarterly Interview Survey,
an interviewer visits every consumer unit in the sample
every 3 months over a 12-month period. It was designed to
obtain data on the types of expenditures respondents can be
expected to recall for a period of 3 months or longer. For
the Diary Survey, consumer units are asked to complete a
record of expenses for two consecutive 1-week periods. It
was designed to obtain detailed data on frequently purchased
small items, such as food and beverages (both at home and in
eating places).
Integrating data from the Interview and Diary Surveys
provides a complete accounting of expenditures and income,
which neither survey component alone is designed to do.
Expenditure levels and expenditure shares (the percent of the
total spent on each category) shown in this report result from
integrating the Diary and Interview Survey data.
to age, sex, race, marital status, and family composition is collected from each survey respondent in a Household Characteristics Questionnaire. The questionnaire also asks for information on work experience, occupation, industry, retirement
status, and income. Income includes member earnings from
wages and salaries, net income from a business or profession,
net income from a farm, and income from all other sources.
Data on household characteristics are collected to determine
the eligibility of the family for inclusion in the population
covered by the Consumer Price Index (CPI), to classify families by family type for purposes of analysis, and to adjust for
nonresponse by families who do not cooperate in the survey.
The data also provide the link between the Diary and Interview Surveys to permit the integration of the data by demographic characteristics.
Quarterly Interview Survey. The quarterly interview portion
of the survey is designed to collect data on major items of
expense, household characteristics, and income. The survey
covers expenditures that one would expect respondents to
recall for 3 months or longer, such as those for property, automobiles, and major appliances, and those that occur on a regular basis, such as rent, insurance premiums, and utilities. The
survey includes detailed data on an estimated 60 to 70 percent
of total household expenditures. In addition, global estimates,
that is, expense patterns for a 3-month period, are obtained for
food and other selected items, accounting for an additional 20
to 25 percent of total expenditures. Each sample household
is interviewed once per quarter, for five consecutive quarters.
Data collected in each quarter are estimated independently,
so annual estimates do not depend upon the participation of a
consumer unit for the full five quarters.
New panels are introduced into the interview sample on a
monthly basis, as other panels complete their participation.
For the Interview Survey as a whole, 20 percent of the sample
is dropped, and a new group added each quarter. This rotating
procedure allows panel estimates to reflect population changes; it also provides operational efficiency by distributing interviewer workload across time.
For the initial interview, information is collected on
demographic and family characteristics and on the inventory of major durable goods of each consumer unit. Expenditure information is also collected in this interview, with a
1-month recall. Expenditure information is used, along with
the inventory information for bounding purposes to minimize
telescoping errors. These double counting errors, common in
Description of survey
BLS contracts with the U.S. Census Bureau to carry out data
collection for both surveys. In the Interview Survey, a U.S.
Census Bureau field representative meets with respondents
and collects expenditure and income data via a computer-assisted personal interview questionnaire. In the Diary Survey,
respondents are asked to report all expenditures made during
their 2-week participation in the survey. Expenditures and related data are recorded in a self-reporting record of daily living expenses. All data collected in both surveys are subject
to confidentiality requirements that prevent the disclosure of
respondents’ identities or such geographic identifiers that may
lead to their identification.
In addition to the Interview Survey questionnaire and the
Diary Survey record of daily expenses, information pertaining
49
retrospective interviews, result from a tendency to report past
events in the reference period of the survey.
The second through fifth interviews use uniform questionnaires to collect expenditure information in each quarter. In
the second and fifth interviews, information also is obtained
on income, such as wage and salary earnings, unemployment
compensation, child support, and alimony, as well as information on the employment of each household member. For new
consumer unit members and members who started work since
the second interview, the interviewers ask for wage, salary,
and other information on employment in the third and fourth
interviews. If there is no new employment information, information is carried over from the second interview to the third
and fourth. In the fifth interview, a supplement is used to collect changes in assets and liabilities.
Households that move away from the sample address between interviews are dropped from the survey. New households that move into the sample address are screened for eligibility and, if found qualified, are included in the survey.
•
Those that are collected solely in the Diary Survey, such
as detailed food expenditures, personal care products,
postage, housekeeping supplies, and nonprescription
drugs.
•
Those that are collected solely in the Interview Survey,
such as expenditures made on trips (overnight and
longer), and information on reimbursed expenditures,
such as for medical care and automobile repairs.
•
Those that are collected in both the Diary and Interview
Surveys, and where a sufficient amount of detail is
collected in both of them. For this group of items, the
survey with more statistically reliable data must be
determined.
In the past, the source selection procedure was run periodically every few years; since 2007, it will be run every 2 years.
The source selection procedure is run separately for each item
category in the third group. The item categories are called
“universal classification codes” (UCCs), which are members
of a 6-digit coding scheme that classifies expenditures. These
are the lowest level at which CE expenditures are tabulated.
The Consumer Expenditure Survey has approximately 900
expenditure UCCs, of which 275 are in the first group, 375
are in the second group, and 250 are in the third group.
For the third group, it is first determined whether there is
a statistically significant difference between the two surveys.
If there is, then the survey with the larger mean annualized
expenditure per consumer unit is selected as the source. If
there is not, the survey currently used as the source continues
to be used.
The procedure was most recently run in 2007, using data
from 2004 to 2006. It started by counting the number of expenditures in the database for each survey to determine whether
credible expenditure estimates could be produced from them.
The threshold for credibility used in the procedure was 60 expenditures per year. If there were 60 or more expenditures
per year in both surveys, and one of the surveys had a significantly higher mean annualized expenditure, then the survey
with the higher mean was selected as the source of data for the
published expenditure estimates. If only one of the surveys
had 60 or more expenditures per year, then the survey with 60
or more expenditures was selected as the source. Otherwise,
the survey currently used as the source was selected again.
The level of statistical significance was determined from
the z-score
xI − xD
,
z=
2
2
SE I + SE D where x– I and x– D are the mean annualized expenditures per
consumer unit in the Interview and Diary surveys, and where
SEI and SED are their standard errors. The z-scores were computed separately for each of the 3 years, and a weighted average of them was computed, with more weight given to the
more recent years. The Interview survey was chosen when the
weighted z-score was greater than 1.645 or when the smallest
Diary Survey. The Diary portion of the survey collects expenditure data for small items purchased on a daily or weekly
basis, such as food, beverages, food consumed away from
home, housekeeping supplies, nonprescription drugs and
medical supplies, and personal care products and services.
However, participants are asked to record all purchases made
each day for two consecutive 1-week periods. Respondents
receive each weekly diary during a separate visit by a Census Bureau interviewer. Data collected in each week are used
independently in the annual estimates, so participation of a
consumer unit for both weeks is not required. However, most
respondents participate for both weeks.
Beginning in 2004, the Consumer Expenditure Survey
(CE) implemented multiple imputation of income to provide
estimated values for missing income data. The process is
applied to both Interview and Diary Survey data. Data from
2004 onward are not strictly comparable to data from previous
years. For an explanation of income imputation, see “Changes
to the 2004 and 2005 Consumer Expenditure Survey published
tables and selected highlights” of the 2004–2005 CE biennial
report.
Source selection
The Diary and Interview Surveys overlap considerably in their
coverage of household expenditures. As mentioned above, the
Interview Survey is designed to obtain data on major expenditures such as automobiles and major appliances that respondents can be expected to recall for 3 months or longer, while
the Diary Survey is designed to obtain data on small frequently purchased items such as food and beverages. When data for
an item category are collected in both surveys, BLS uses the
data from only one of the surveys in its published expenditure
estimates. The survey with more statistically reliable data is
chosen. The process of choosing the survey with more statistically reliable data is called “source selection.”
For source selection, the item categories are divided into
three groups:
50
For the Interview Survey, approximately 14,000 addresses
are contacted in each calendar quarter of the year. Onefifth of the addresses contacted each quarter are new to the
survey and provide “bounding” interviews that provide
baseline data, which are not used in the survey’s published
expenditure estimates. Excluding these bounding interviews
and interviews not completed due to refusals, vacancies,
ineligibility, or the nonexistence of a housing unit at the
selected address, usable interviews are obtained from
approximately 7,000 households each quarter. After a housing
unit has been in the sample for five consecutive quarters, it is
dropped from the survey and a new housing unit is selected
to replace it.
of the three yearly z-scores was greater than 1.000. The Diary
survey was chosen when the weighted z-score was less than
-1.645 or when the largest of the three yearly z-scores was less
than -1.000.
Sample design
The Consumer Expenditure Survey is a nationwide household
survey designed to measure the expenditures of the U.S.
civilian noninstitutional population, which comprises about
98 percent of the total U.S. population. CE includes people
living in houses, condominiums, apartments, and group
quarters, such as college dormitories. It excludes military
personnel living on base, nursing home residents, and people
in prisons.
The selection of households for the survey begins with the
definition and selection of geographic areas called “primary
sampling units” (PSUs). PSUs are small groups of counties
that have been grouped together into geographic entities
called “core-based statistical areas” (CBSAs). The CBSAs
are categorized into three groups based on the population of
the largest urbanized area inside them: Metropolitan Statistical Areas, which are CBSAs whose largest urbanized area
has more than 50,000 people; Micropolitan Statistical Areas,
which are CBSAs whose largest urbanized area has between
10,000 and 50,000 people; and non-CBSA areas, which are all
remaining areas.
The sample of PSUs used in the 2006–2007 survey consists of 91 PSUs, of which 75 urban PSUs are also used by the
Consumer Price Index program. The 91 PSUs are classified
into four categories:
Response rates
Response data for the 2006–2007 Interview and Diary Surveys are shown in tables B-1 and B-2. For the Interview Survey, totals refer to housing units in the second through fifth
quarters of the survey (the non-bounding interviews), with
each unique housing unit providing up to four usable interviews. For the Diary Survey, the totals refer to housing units
in weeks 1 and 2 of the survey, with each unique housing unit
providing up to two usable interviews. Most Diary respondents participate for both weeks.
There are three general categories of nonresponse:
• Type A nonresponses are refusals, temporary absences,
and noncontacts.
• Type B nonresponses are vacant housing units, housing
units with temporary residents, and housing units under
construction.
• 21 “A” PSUs, which are metropolitan CBSAs with a
population over 2.7 million people.
• 38 “X” PSUs, which are metropolitan CBSAs with a
population under 2.7 million people.
• Type C nonresponses are nonexistent housing units,
destroyed or abandoned housing units, and housing
units converted to nonresidential use.
• 16 “Y” PSUs, which are micropolitan CBSAs.
• 16 “Z” PSUs, which are non-CBSA areas, and are often
referred to as “rural” PSUs.
Type A nonresponses are considered to be in-scope and
eligible units, because they were able to participate in
the survey but either chose not to do so or could not be
contacted. Type B nonresponses are considered to be
in-scope but ineligible units, because the addresses are vacant. Type C nonresponses are considered to be out-of-scope
units.
Response rates are defined to be the percent of eligible
housing units (i.e., the designated sample less Type B and
Type C nonresponses) from which usable interviews are collected. In the 2007 Interview Survey, there were 37,016 eligible housing units, from which 27,335 usable interviews were
collected, resulting in a response rate of 73.8 percent. In the
2007 Diary Survey, there were 19,595 eligible housing units,
from which 13,747 usable interviews were collected, resulting
in a response rate of 70.2 percent.
Within these 91 PSUs, the sampling frame (the list of
addresses from which the sample is drawn) is generated from
the 2000 Census 100-percent detail file. The sampling frame
is augmented by new construction permits and extra housing
units identified through coverage improvement techniques to
compensate for recognized deficiencies in the file.
From this sampling frame, the U.S. Census Bureau selects
a representative sample of approximately 12,000 addresses
per year to participate in the Diary Survey. Usable diaries
are obtained from approximately 7,000 households at those
addresses. Diaries are not obtained from the other addresses
due to refusals, vacancies, ineligibility, or the nonexistence of
a housing unit at the selected address. The actual placement of
diaries is spread equally over all 52 weeks of the year.
51
quarterly. Each CU is given its own unique calibration factor.
There are infinitely many sets of calibration factors that make
the weights add up to the 24 known population counts, and
the set that is used minimizes the amount of change made to
the “initial weights” (initial weight = base weight x weighting
control factor x noninterview adjustment factor). The same
population controls are used for both the Interview and Diary
Surveys, hence both surveys have the same demographic
estimates for the controlled variables.
Table B-1. Analysis of response in the 2006--07 Interview
Survey
Sample unit
2006
2007
Housing units designated for the
survey...................................................
46,789
45,996
Less: Type B and C nonresponses....
9,080
8,980
Equals: Eligible units.............................
37,709
37,016
Less: Type A nonresponses...............
8,842
9,681
Equals: Interview units..........................
28,867
27,335
Percent of eligible units interviewed.......
76.6
73.8
Data collection and processing
Due to differences in format and design, the Interview Survey
and the Diary Survey are collected and processed separately.
The U.S. Census Bureau, under contract with BLS, carries out
data collection for both. In addition to its collection duties,
the Census Bureau does field editing and coding, checks
consistency, ensures quality control, and transmits the data to
BLS. In preparing the data for analysis and publication, BLS
performs additional review and editing procedures.
Table B-2. Analysis of response in the 2006--07 Diary
Survey
Sample unit
2006
2007
Housing units designated for the
survey...................................................
24,320
24,642
Less: Type B and C nonresponses....
4,844
5,047
Equals: Eligible units.............................
19,476
19,595
Less: Type A nonresponses...............
5,021
5,848
Equals: Interview units..........................
14,455
13,747
Percent of eligible units interviewed.......
74.2
70.2
Quarterly Interview Survey. Beginning April 2003, Census
Field Representatives (FRs) began collecting the Interview
data using a Computer Assisted Personal Interview (CAPI)
instrument. This was a major improvement from the paper
and pencil data collection that had been in place since 1980.
The CAPI instrument enforces question skip patterns, allows
for data confirmation of high expenditure values, and reduces
processing time. The FR performs some coding of expenses—
by selecting from a predetermined list—for vehicle make and
model, trip destination, and job types for alterations, maintenance, and repair.
Data are electronically transferred from the FR’s laptop
at completion of the interview to the Census Master Control
System. The Census Bureau Demographics Surveys Division
then reformats the data into datasets and does special processing for output to BLS (such as converting missing values to
special characters and merging data records into the required
BLS output structure.) Some data, like vehicle and mortgage
records, are copied into an input file that is loaded on the laptops for subsequent interviews the next quarter. This way, a
few fields are updated each quarter, rather than recollecting
the entire data record.
At BLS, a series of automated edits are applied to monthly
data. These edits check for inconsistencies, identify missing
expenditure amounts for later imputation, impute missing
demographic variables, calculate weights, and adjust data to
include sales tax and to exclude business expenses or reimbursed expenditures.
Monthly data files are then combined into quarterly databases, and a more extensive data review is carried out. This
step includes a review of the following: counts and means by
region, family relationship coding inconsistencies, and selected extreme values for expenditure and income categories.
Other adjustments convert mortgage and vehicle payments
into principal and interest (using associated data on the interest rate and term of the loan). In addition, BLS verifies the
Weighting
Each consumer unit (CU) in the Consumer Expenditure Survey
represents a given number of similar CUs in the U.S. civilian
noninstitutional population. The translation of sample CUs
into the universe of CUs is known as “weighting.” Several
factors are involved in computing the weight of each CU for
which a usable interview is obtained. Each CU is initially
assigned a base weight, which is equal to the inverse of its
probability of being selected for the sample. Base weights in
the Consumer Expenditure Survey are typically around 10,000,
which means that a CU in the sample represents 10,000 CUs
in the U.S. civilian noninstitutional population—itself plus
9,999 other CUs that were not selected for the sample. The
base weight is then adjusted by the following factors:
Weighting control factor. This adjusts for subsampling in the
field. Subsampling occurs when a field representative visits a
particular address and discovers multiple housing units, where
only one housing unit was expected.
Noninterview adjustment factor. This adjusts for interviews
that cannot be conducted in occupied housing units due to a
CU’s refusal to participate in the survey or the fact that no one
is home. This adjustment is based on the region of the country
(Northeast, Midwest, South, West), household tenure (owner/
renter), CU size, and race of the CU’s reference person.
Calibration factor. This adjusts the weights to 24 “known”
population counts to account for frame undercoverage. These
known population counts are for age, race, household tenure,
region of the country, and urban/rural. The population counts
come from the Current Population Survey, and they are updated
52
various data transformations it performs. Cases of questionable data values or relationships are investigated, and errors
are corrected prior to release of the data for public use.
Three major types of data adjustment routines—imputation,
allocation, and time adjustment—improve estimates derived
from the Interview Survey. Data imputation routines account
for missing or invalid entries and affect all fields in the database,
except assets. Missing or invalid attributes or expenditures are
imputed. Allocation routines are applied when respondents
provide insufficient detail to meet tabulation requirements.
For example, combined expenditures for the fuels and utilities
group are allocated among the components of that group, such
as gas and electricity. Time adjustment routines are used to classify expenditures by month, prior to aggregation of the data to
calendar-year expenditures. Tabulations are made before and
after the data adjustment routines to analyze the results.
The interview and diary surveys implemented multiple imputations of income data, beginning with the publication of
the 2004 data. Prior to that, only income data collected from
complete income reporters were published. However, even
complete income reporters did not provide information on
all sources of income for which they reported receipt. With
the collection of bracketed income data starting in 2001, this
problem was reduced but not eliminated. A limitation was that
bracketed data only provide a range in which income falls,
rather than a precise value for that income. In contrast, imputation allows income values to be estimated when they are not
reported. In multiple imputations, several estimates are made
for the same consumer unit, and the average of these estimates
is used in the published data.
hold and consumer unit codes are assigned to each household
member, and industry and occupation codes are entered for
each working member. After an initial clerical screening, data
are key-entered into electronic formats and a computer file of
the database containing these data is produced and transmitted monthly to BLS, along with image files of questionnaires.
Data then are processed to calculate population weights
based on BLS specifications, impute demographic characteristics for missing or inconsistent demographic data, impute
values for weeks worked when nonresponse is encountered,
and apply appropriate sales taxes to the expenditure items.
Using three monthly diary data files, BLS creates a quarterly database and screens it for invalid coding and inconsistent
relationships, as well as for extreme values recorded or keyed
erroneously. BLS then corrects any coding and extreme-value
errors found.
Two types of data adjustment routines—allocation and imputation—improve the Diary Survey estimates. Allocation routines
transform reports of nonspecific items into specific ones. For
example, when respondents report expenditures for meat rather
than beef or pork, allocations are made, using proportions derived
from item-specific reports in other completed diaries. BLS imputes missing attributes, such as age or sex or package type, needed for mapping Diary expenditures. Income data from the Diary
Survey are processed in the same way as in the Interview Survey.
Reliability of the data
Sample surveys are subject to two types of errors, sampling
and nonsampling. Sampling error is the uncertainty in the
Consumer Expenditure Survey estimates caused by the fact
that data are collected from a sample of consumer units across
the United States, instead of collecting data from every consumer unit. (The United States has approximately 120 million consumer units.) In 2007, usable data were collected
in 27,335 quarterly interviews in the Interview Survey, and
in 13,747 weekly diaries in the Diary Survey. Non-sampling
error is the rest of the error. Non-sampling error includes
things like incorrect information given by respondents, data
processing errors, and so on. Non-sampling error occurs, regardless of whether data are collected from a sample of consumer units, or from the complete universe of consumer units.
The most common measure of the variability in a survey’s
estimates caused by sampling error is the standard error of the
estimates—the square root of the variance. The standard error
of the Consumer Expenditure Survey estimates can be used to
construct confidence intervals to test various statistical hypotheses. Tables showing the standard errors of the Consumer Expenditure Survey estimates are available on the Internet at the
Consumer Expenditure Survey webpage: www.bls.gov/cex.
The Bureau of Labor Statistics is constantly working to
reduce error in the Consumer Expenditure Survey. Sampling
error is reduced by using a sample of consumer units that is
as large as possible given resource constraints. Non-sampling
error is reduced through a series of computerized and professional data reviews, through continuous survey process improvements, as well as through theoretical research.
Diary Survey. At the beginning of the 2-week collection period, the Census Bureau interviewer, using the Household
Characteristics Questionnaire (a CAPI instrument), records
demographic information on members of each sampled consumer unit. At this time, the interviewer also leaves the Diary
questionnaire—or daily expenditure record—with the consumer unit, to record expenditures for the week.
Respondents record all expenses incurred during their participation in the survey in the diary questionnaire, a self-reporting, product-oriented diary. The diary is divided by day of
purchase and by a broad classification of goods and services.
At the end of the first week, the interviewer collects the
diary, reviews the entries, answers any questions, and leaves
a second diary. The interviewer picks up the second diary at
the end of the second week and reviews the entries. During
this time, the interviewer again uses the Household Characteristics Questionnaire to collect previous-year information on
work experience and income. Each week of a consumer unit’s
participation in the survey is treated as a separate occurrence.
The Census Bureau performs preliminary processing activities, including a number of data edits and adjustments. Data
in the diaries are reviewed during a field edit for completeness
and consistency. All notes are reviewed, so expenditure data
can be transcribed to the questionnaire for keying. In addition,
item codes are assigned to reported expenditure items, house53